OptFedAvg: A Client Selection Optimizer for Efficient Federated Learning

Federated Machine Learning (FML) has emerged as a solution for collaboratively training artificial intelligence models while improving data privacy. Unlike traditional approaches, FML allows multiple edge devices to train models locally without sharing sensitive data, making it especially useful in environ-ments where privacy and security are essential. However, the performance of the global model largely depends on the quality of the local models and how they are aggregated. Optimizing the membership of participating devices is key to ensuring good convergence of the global model. For this reason, this work introduces two optimizers designed to select the best clients based on the accuracy of their local models, training times, and energy consumption. The selection process relies on both pre-tabulated data for each device type and dynamic metrics collected throughout the local training rounds. This approach contrasts with many state of the art methods, where client selection at the beginning is typically random. In our case, the focus is on accelerating the training process by leveraging pre-tabulated performance metrics from the devices, collected in similar classification tasks, to make more informed decisions from the start. The results obtained using this strategy show promising results in both training efficiency and overall model performance.